rocWMMA support

[radudiaconu0]

Could you implement rocWMMA to use with Navi3 GPU? From what i understood it uses the AI accelerotors present in them for faster matrix multilplication. I guess this could make it faster for DL tasks.

https://github.com/ROCmSoftwarePlatform/rocWMMA

[vchuravy]

One could, but it would require probably quite a bit of work by someone.

CUDA.jl WMMA support was the work of a full-time master student.

[radudiaconu0]

One could, but it would require probably quite a bit of work by someone.

CUDA.jl WMMA support was the work of a full-time master student.

wekk it has to be done at some point. why only nvidia to get all the goodies? :P

[pxl-th]

For matrix multiplication we are using rocBLAS, adding wmma support won't affect its performance.

[pxl-th]

And at this moment matrix multiplication is not a bottleneck in DL applications for AMDGPU.
Timely memory freeing is.

[vchuravy]

wekk it has to be done at some point. why only nvidia to get all the goodies? :P

Are you volunteering?

[radudiaconu0]

wekk it has to be done at some point. why only nvidia to get all the goodies? :P

Are you volunteering?

i would like to try